A watermark in the corner of a photo is the easiest kind to lose. We put a “© Your Name” mark 15% of the image width wide in the bottom-right corner of three photos, then cropped each one to three common shapes: square, 4:5 portrait and 16:9 wide. The square crop kept 11% of the mark. The 4:5 and 16:9 crops kept none of it, and nobody had to paint anything out.
That is the whole problem with a corner mark, and it is why we recommend a repeated mark across the picture if copying worries you. The numbers below are from our own Add watermark to images tool, measured on 24 September 2026 in Chromium 153.
How the test worked
Three public-domain photos from Wikimedia Commons, all 3840 by 2560 pixels: a mountain landscape from Grand Teton National Park, a NASA crew portrait and a NASA control room. Each got the same white text in Montserrat Extra Bold with the tool’s thin dark edge, in six layouts: one mark in the corner at 15% and 30% of the width, one in the centre at 30%, a straight repeat at 30%, and a diagonal repeat at 15% and 30%. Each was saved as the tool saves a JPG, at quality 90. We then measured three things: how much of the picture the mark covers, how much of the mark survives a centre crop, and what is left after painting the mark out with the two inpainting methods built into OpenCV 5.0.
What a crop leaves
| Layout | Covers | Square crop | 4:5 crop | 16:9 crop | Largest clean area after a 4:5 crop |
|---|---|---|---|---|---|
| Corner, 15% wide | 0.26% | 11% of the mark | None | None | 100% |
| Corner, 30% wide | 1.0% | 49% | 32% | 26% | 90% |
| Centre, 30% wide | 1.0% | 100% | 100% | 100% | 48% |
| Repeated, 30% wide | 13.7% | 66% | 52% | 74% | 11% |
| Diagonal repeat, 15% wide | 13.2% | 68% | 55% | 85% | 2% |
| Diagonal repeat, 30% wide | 11.9% | 67% | 51% | 93% | 7.5% |
The last column is the one to read. It is the biggest rectangle of the cropped photo that doesn’t touch the mark, which is what someone could cut out and use as if it were unmarked. A corner mark leaves 90 to 100% of the photo clean. A centre mark survives every crop but still leaves about half of the picture clean on one side of it. The diagonal repeat at 15% left no clean area bigger than 2% of the photo.
The price is plain: a repeated mark covers 12 to 14% of the picture, against 1% or less for a single mark. People will see it. That is the trade, and for proofs you send to a client before they pay it is the right one.
Painting it out
Cropping doesn’t work on a mark in the middle, so the next step is inpainting: filling the marked pixels from what’s around them. We gave OpenCV’s Telea and Navier-Stokes inpainting the exact shape of our mark, which is more than anyone copying your photo would have, and compared the result with the original photo.
| Layout | Landscape, dB | Crew portrait, dB | Control room, dB | Whole-photo similarity after |
|---|---|---|---|---|
| Corner, 15% wide | 20.0 | 31.4 | 25.2 | 0.989 to 0.995 |
| Corner, 30% wide | 19.7 | 25.1 | 19.7 | 0.982 to 0.994 |
| Centre, 30% wide | 17.5 | 24.2 | 17.5 | 0.987 to 0.991 |
| Repeated, 30% wide | 19.1 | 21.0 | 18.2 | 0.928 to 0.956 |
| Diagonal repeat, 30% wide | 20.2 | 19.2 | 17.4 | 0.941 to 0.955 |
Inside the painted area, no layout came back clean: 17 to 31 dB is a smear you can see when you look for it. On the crew portrait the corner mark came out best, because it sat on a plain blue flight suit; the clearest giveaway was a blob where the mark had crossed a pencil. The repeated marks sat on faces, and inpainting a line of text across a face left the face visibly smudged. Across a whole photo, the repeated layouts dropped similarity to 0.93 to 0.96, while the corner and centre marks stayed above 0.98, because their damage is 1% of the picture or less.
Only simple inpainting went into this. Generative fill tools work differently, inventing what was under the mark from what photos usually look like, and we haven’t measured any of them. Don’t read our smears as the best anyone can do.
See-through marks can be subtracted
A mark at 60% opacity lets 40% of the photo through, and a blend like that can be undone. If you know the mark’s exact colour, shape and opacity, each pixel of the original is the watermarked pixel minus the mark’s share, divided by what was left. We tried it on our own saved JPGs, knowing the mark exactly. Inside the mark the match to the original went from 8 to 12 dB before to 27 to 35 dB after, clean enough that the diagonal mark across the crew’s faces was hard to find.
Nobody copying your photo knows your mark that exactly, but they don’t need to. Google researchers showed in 2017 that when the same mark sits in the same place on many photos, it can be estimated from the collection and then removed from each one, and they suggested varying it slightly per photo to stop that. A fully opaque mark can’t be subtracted at all: nothing of the photo is left under it, so it can only be painted over.
What we’d do
For photos you publish and want credited, a small corner mark is fine; it is a signature, and it will be cropped off by anyone who wants it gone. For proofs and previews you don’t want used, a diagonal repeat at 15 to 30% of the width, at high opacity, left the smallest clean area in every crop we tried and forced any painting-out across faces and detail. Our tool’s default is 60% opacity; raise it for anything that matters.
Sources
- Wikimedia Commons: Snake River Overlook (Grand Teton NPS, public domain)
- Wikimedia Commons: STS-135 crew portrait (NASA, public domain)
- Wikimedia Commons: Orion Mission Evaluation Room team (NASA, public domain)
- Google Research: Making visible watermarks more effective (CVPR 2017 paper summary)
- OpenCV source: cv::inpaint and its Telea and Navier-Stokes methods